SSAD 2026: Diffusion models and public datasets both fail to generalize
abursuc · x · 2026-09-18
Pessimistic field notes from SSAD 2026: diffusion-based synthetic data on point clouds yields only a few percentage points of improvement; public datasets are insufficient, open-loop and novel-view-synthesis setups fail to generalize, and end-to-end pipelines still need a babysitter. Overall, the generalization of current autonomous-driving/synthetic-data approaches appears overstated.
More from Research
- SSAD 2026 closing talk: Dirk Hornung on OpenXLA as a ready-to-use end-to-end compiler stack — abursuc · 2026-09-18
- VākQA: A Telugu spoken QA benchmark exposing flaws in LLM-as-judge evaluation — SPL-IIITH · 2026-09-18
- Co-Designing AI With Older Adults Boosts Adoption, Usability, WSU Study Finds — pshrink · 2026-09-18
- Salesforce's DarwinX uses an evolutionary approach to fix self-optimizing agent harnesses — bendee983 · 2026-09-18
- tokenbender: no benchmark can capture frontier models' inhuman blind spots in SWE/MLE — tokenbender · 2026-09-18
- Chaining Yes/No Judgments in jev Yields an Embedding-Like Vector — zsakib_ · 2026-09-18